2026
EMNLP
EMNLP 2026
How many raters do we need to recognize a bad translation via perplexity?
Abstract
AbstractWe evaluated candidate translation pairs with various kinds of damage using perplexity assigned to them by LLMs to see how it holds up as a fluency and adequacy metric. We also explored how the discriminating power of perplexity can be improved by ensembling.